# Sub-agents

> How RoGenie splits big jobs across specialist sub-agents that work in parallel, which model each role runs on, and how to stop or resume them.

For a big job with independent parts, RoGenie works like a team lead. The main agent, the **orchestrator**, plans the work and hands separate parts to **sub-agents** that build at the same time. Three themed zones, a shop, an inventory and a quest system can all be built in parallel, which is much faster than one step at a time.

## The roles

Each sub-agent gets a specialist role with its own instructions:

| Role | Typical work |
|---|---|
| **Builder** | Maps, buildings, props and geometry |
| **Scripter** | Game systems and Luau code |
| **Debugger** | Finding and fixing bugs |
| **UI Designer** | GUIs, menus and HUDs |
| **VFX Artist** | Particles, beams, lighting and effects |
| **Asset & Audio Finder** | Toolbox models, audio and icons |
| **Animator** | Animations and rigs, in Studio or in [Blender](https://rogenie.ai/docs/blender) when it is connected |
| **3D modelling (Blender)** | Models, sculpts, UVs and materials in your [Blender](https://rogenie.ai/docs/blender), imports the result into Studio, and renders icons and thumbnails |
| **Verifier** | Checking work and running reviews |
| **Researcher** | Looking things up in the Roblox docs and RoGenie's knowledge |

Sub-agents work in their own isolated context and are given separate parts of your place to write to, so they do not overwrite each other. RoGenie also uses sub-agents for parallel investigation, such as auditing many scripts or searching for several assets at once.

### The Blender sub-agents

The **3D modelling (Blender)** sub-agent works in your own Blender through the [RoGenie Bridge](https://rogenie.ai/docs/bridge). It models the asset in its own collection, checks it in Blender's viewport, sends it into your place as real MeshParts, confirms the import landed in Studio, and then reports back. If any part did not come across, the report says which. Its import has to name the objects it sends, so your own objects in the scene are not exported with it. The same agent also makes game icons and thumbnails: Blender renders the image, and the agent sets up the shot and checks the result. Uploading a render needs your approval, so it hands the render to the main agent, which shows you the approval card (unless you chose **Always allow** for uploads in that chat).

This sub-agent works only while Blender is connected and allowed, so open Blender before you ask for Blender work. While Blender is closed it is not available, and if it can't start, RoGenie tells you what is missing. Only one sub-agent uses Blender at a time, because everything happens in one Blender scene. A second one shows **Waiting for Blender** until the first finishes or has not used Blender for 10 minutes; after an hour of waiting it pauses and can be resumed. The **Animator** also uses Blender when it is connected, for character animations and rigged models, and works in Studio when it is not.

## Following them

While sub-agents run, a **Parallel sub-agents** panel in the work log shows each one with its role, progress and result, and a summary above the message box tracks them all ("4 sub-agents completed — 36 steps, 18 actions"). When they finish, the orchestrator reads their results, puts the pieces together and verifies the whole.

## Choosing a model for each role
Open **Settings › Agent › Agent & sub-agents** (also reachable from the model picker). There you can pick:

- the **Orchestrator model**, which plans the work and assigns the sub-agents, and
- a model for each **specialist sub-agent** role.

**Same as chat** runs that role on the model you picked for the chat. By default, the **Debugger**, **Asset & Audio Finder**, **Verifier** and **Researcher** run on **Auto (intelligent & cheap)**, and the other roles, the **Animator** and **3D modelling (Blender)** included, follow the chat's model. The 3D modelling agent's icon work runs on the model you pick for it, like the rest of its job. A good pattern is a cheap model for asset searches and research and a strong one for scripting.

[Code reviews](https://rogenie.ai/docs/code-review) run on the Verifier's model.

## Stop, resume and wait

The orchestrator has tools to manage its sub-agents, and you can ask for them in plain words:

- **Stop Sub-Agents** stops one or all of the sub-agents a run spawned.
- **Resume Sub-Agents** brings back sub-agents that paused or failed, for example because a model route went down or a rate limit hit. A resumed sub-agent keeps the work it already applied.
- **Wait for Sub-Agents** makes the orchestrator wait for some or all of them before it continues.

Pressing **Stop** in the app stops the whole run, sub-agents included.

## Cost

Every sub-agent step is a model call billed to your [quota](https://rogenie.ai/docs/usage-and-quota), at the multiplier of the model its role runs on. Parallel work is faster, not cheaper: picking cheap models for simple roles is what saves quota.

## Turning sub-agents off

Switch off **Sub-Agents** in [Tool access](https://rogenie.ai/docs/tool-access) to make RoGenie do all the work in a single agent for the session.